A bearing composite fault intelligent diagnosis method based on multi-view feature dynamic fusion

Abstract

The application discloses a bearing composite fault intelligent diagnosis method based on multi-view feature dynamic fusion. The technical steps of the method are as follows: (1) multi-view feature representation: the original data is processed by using fast Fourier transform and continuous wavelet transform to obtain frequency domain data and time-frequency image; (2) feature extraction: a CNN-BiLSTM-Attention-KAN and ResNet-KAN model are constructed to extract frequency domain features and time-frequency features respectively; (3) diagnosis model construction: the feature fusion is realized by using a two-stage dynamic fusion framework of fusion uncertainty perception and adaptive learning, and a binary cross-entropy loss function based on an adaptive adjustment mechanism is used to train the model in the first stage; (4) model optimization: the model parameters are optimized by using an Adam algorithm. Through the multi-view feature dynamic fusion, the application effectively alleviates the problem of insufficient generalization ability caused by the fact that a traditional diagnosis method only depends on single-view features, so that the diagnosis precision is further improved, and the application has high practical application value.

Description

Technical Field

[0001] This invention belongs to the field of bearing composite fault diagnosis, specifically involving an intelligent diagnosis method for bearing composite faults based on dynamic fusion of multi-view features. Background Technology

[0002] With the continuous growth in demand from the high-end equipment manufacturing industry, bearings, as key components of rotating machinery, have been widely used in industrial fields such as intelligent manufacturing and aerospace. Their reliability and stability are crucial to ensuring production efficiency and operational safety.

[0003] However, in practical applications, bearings become particularly prone to failure in rotating machinery due to the combined effects of prolonged operation, harsh conditions, material aging, and insufficient maintenance. Timely and accurate detection and classification of these faults are crucial for ensuring normal equipment operation; otherwise, undetected faults can lead to decreased performance, unexpected shutdowns, and even safety accidents. Previous studies on bearing complex fault diagnosis largely relied on single-view feature representations, which struggled to comprehensively capture the multidimensional information of complex faults and lacked theory-driven multi-view feature fusion strategies, thus failing to meet the demands for high-precision diagnosis. Based on these considerations, this invention proposes an intelligent bearing complex fault diagnosis method based on multi-view feature dynamic fusion, integrating multi-view feature representation and dynamic fusion mechanisms. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic method for bearing composite faults based on dynamic fusion of multi-perspective features. This method first uses Fast Fourier Transform and Continuous Wavelet Transform to represent the multi-perspective features of the original data. Then, deep learning models are constructed for feature extraction from frequency domain data and time-frequency images respectively. Finally, a two-stage dynamic fusion framework integrating uncertainty perception and adaptive learning strategies is employed to achieve feature fusion: In the first stage, an uncertainty perception method is used to generate dynamic weights for fusing feature vectors, and a binary cross-entropy loss function based on an adaptive adjustment mechanism is introduced. The loss weights are adaptively allocated according to the sample difficulty to optimize the quality of the fused features and improve the model's robustness to noise. In the second stage, a regression model is used to further adaptively learn the optimal weight allocation among the multi-perspective features to compensate for the bias in uncertainty estimation, thereby improving the robustness and predictive performance of the intelligent diagnostic method for bearing composite faults.

[0005] A method for intelligent diagnosis of composite bearing faults based on dynamic fusion of multi-view features includes the following steps:

[0006] S100: Obtain raw data on composite bearing faults and construct a multi-label dataset based on bearing fault types;

[0007] S200: Through Fast Fourier Transform and Continuous Wavelet Transform, the original data is converted into frequency domain data and time-frequency image respectively, so as to realize the multi-view feature representation of the original data;

[0008] S300: CNN-BiLSTM-Attention-KAN and ResNet-KAN models were constructed for frequency domain data and time-frequency images, respectively, to extract frequency domain features and time-frequency features;

[0009] S400: For the extracted frequency domain features and time-frequency features, a two-stage dynamic fusion framework is used for feature fusion. In the first stage, dynamic weights are generated through uncertainty perception and the model is trained using a binary cross-entropy loss function based on an adaptive adjustment mechanism. In the second stage, adaptive learning is performed through a regression network to further optimize the weight allocation, so as to improve the quality of the fused features and classification performance in the fault diagnosis model.

[0010] S500: Trainable parameters of a fault diagnosis model optimized based on the Adam algorithm;

[0011] S600: Uses the trained model to perform fault diagnosis and obtain the final prediction result.

[0012] In step S200, the calculation process of the Fast Fourier Transform specifically adopts the following formula:

[0013] (1)

[0014] in, It is a discrete-time signal Length, It is the first signal in the frequency domain One frequency component, These are the basis functions of the discrete Fourier transform used to adjust the phase of the frequency components. It is the Fast Fourier Transform function;

[0015] The calculation process of the continuous wavelet transform specifically adopts the following calculation formula:

[0016] (2)

[0017] in, It is the original signal. Morley wavelets are used as basis functions. It is dimensionless time. It is a scaling factor. It is the translation factor.

[0018] In step S300, the convolution operation of the convolutional neural network specifically adopts the following formula:

[0019] (3)

[0020] in, yes The first layer One element, It is the first Convolutional regions of feature maps, It is a weight matrix. It's a bias. It is an activation function;

[0021] The average pooling operation of the convolutional neural network specifically adopts the following formula:

[0022] (4)

[0023] The max pooling operation of the convolutional neural network described above uses the following formula:

[0024] (5)

[0025] in, These are the corresponding elements of the output feature map. For window size, Step size, and All are indexes. It is the corresponding element in the input;

[0026] The calculation process of the BiLSTM model specifically adopts the following formula:

[0027] (6)

[0028] (7)

[0029] (8)

[0030] in, This refers to the LSTM computation process. The input weight matrix for the forward LSTM is... Here is the hidden state weight matrix of the forward LSTM. yes The input vector at time t, For positive LSTM in The hidden state at all times For positive LSTM in The hidden state of time, The input weight matrix for the backward LSTM The hidden state weight matrix of the backward LSTM. yes The input vector at time t, For backward LSTM The hidden state at all times For backward LSTM in The hidden state of time, It is the fusion weight of the forward hidden state. It is the fusion weight of the backward hidden state. It is an activation function;

[0031] The calculation process of the self-attention mechanism specifically adopts the following formula:

[0032] (9)

[0033] in, , It is the sequence length. It is the embedded dimension. , and The following are the weight matrices for trainable queries, keys, and values, respectively. It is the result of calculations using the self-attention mechanism;

[0034] The calculation process of the KAN model specifically adopts the following formula:

[0035] (10)

[0036] in, It is the dimension of the input vector. It is the input vector. It is a one-dimensional continuous function. It is a one-dimensional continuous function within the closed interval [0,1]. and From 0 to Outer sum index from 1 to The inner summation index.

[0037] In step S400, the late-stage dynamic weighted fusion of the two-stage dynamic fusion framework in the multi-view feature fusion specifically adopts the following formula:

[0038] (11)

[0039] in, It is a multi-view classification model. It is the first Feature extraction model from multiple perspectives and The learnable parameters of the multi-view classification model are, in order, the first and second. Learnable parameters of a feature extraction model from multiple perspectives It is the first The fusion weights of features from each perspective , It is the number of viewpoint features. No. Input feature vectors from multiple perspectives;

[0040] The weight design principle of the two-stage dynamic fusion framework adopts the following formula:

[0041] (12)

[0042] in, It is the first The classification accuracy metric of the feature extraction model from the nth perspective, when the nth perspective is... From each perspective, the feature extraction model outputs 1 when it correctly classifies a sample, and outputs a smaller value when it misclassifies a sample. , and The order is the first Predicted and actual labels of samples from different perspectives;

[0043] The theoretical design principles of the two-stage dynamic fusion framework are specifically based on the following formula:

[0044] (13)

[0045] in, It is the accuracy of multi-view classification;

[0046] The output of the predicted probability in the first stage of the two-stage dynamic fusion framework specifically adopts the following formula:

[0047] (14)

[0048] (15)

[0049] (16)

[0050] (17)

[0051] (18)

[0052] (19)

[0053] (20)

[0054] in, It is the predicted probability of TOP-1. It is the predicted probability of TOP-2. It is the first In the feature extraction model from multiple perspectives, the evidence vectors are generated by evidence deep learning, and Softplus is the activation function. It is the first Feature extraction models from various perspectives for categories The predicted probability, , It is the first Feature extraction models from various perspectives for categories The Dirichlet distribution parameters, It is the first Feature extraction models from various perspectives for categories The vector of evidence, ,make ,but , It is the dimension of the labels in the sample. and The order is the first Predicted probabilities of TOP-1 and TOP-2 categories in a feature extraction model from multiple perspectives and The order is the first Dirichlet distribution parameters of the top-1 and top-2 classes in the feature extraction model from each perspective and The order is the first Evidence vectors of the top-1 and top-2 categories in the feature extraction model from each perspective It is the first Predictive uncertainty of feature extraction models from various perspectives It is the first Predictive determinism of feature extraction models from various perspectives Is assigned to the first Dynamic fusion weights of features from various perspectives It is a category The Middle Evidence vectors from feature extraction models at various perspectives It is a category The vector of evidence, It is the first stage of the two-stage dynamic fusion framework for categories. The predicted probability;

[0055] The binary cross-entropy loss function based on the adaptive adjustment mechanism in the first stage of the two-stage dynamic fusion framework adopts the following formula:

[0056] (twenty one)

[0057] (twenty two)

[0058] in, It is the sample label dimension. It is the first The first sample The real label of each tag It is the first The first sample Predicted labels for each label, It is the first The first sample The prediction confidence of each label, It is the first The first sample The predicted probability of each label. It is a hyperparameter that determines whether a sample is easily classified when the model's prediction is correct. Set as When the model's prediction is incorrect, the sample is considered a hard-to-classify sample. Set as ;

[0059] The overall objective function in the first stage of the two-stage dynamic fusion framework is specifically adopted by the following formula:

[0060] (twenty three)

[0061] in, Input data The classification results of the multi-view classification model. It is the first Classification results of a feature extraction model from multiple perspectives;

[0062] The regression-based multi-view dynamic fusion in the second stage of the two-stage dynamic fusion framework specifically adopts the following formula:

[0063] (twenty four)

[0064] (25)

[0065] in, The first one obtained after normalization The fusion weights of features from different perspectives It's about returning to the internet. These are the parameters of the regression network in the second stage. These are the learnable model parameters in the first stage and are frozen in the second stage;

[0066] The objective function in the second stage of the two-stage dynamic fusion framework is specifically calculated using the following formula:

[0067] (26)

[0068] The activation function of the two-stage dynamic fusion framework The specific calculation formula is as follows:

[0069] (27)

[0070] in, It is the model's raw output;

[0071] In step S500, the calculation process for updating the trainable parameters of the model using the Adam algorithm specifically adopts the following formula:

[0072] (28)

[0073] (29)

[0074] (30)

[0075] in, It is the gradient of the parameters. It is the exponentially weighted average of the squared gradients. It is the exponentially weighted average of the gradient. and It is the weighted average of two exponents. and The attenuation coefficient, It is a parameter at time t. It's the learning rate. It is a constant with a very small value. These are the parameters at time t+1, i.e., the updated parameters.

[0076] Compared with existing methods for diagnosing composite bearing faults, the advantages of this invention are as follows: This invention constructs a multi-view feature representation based on frequency domain data and time-frequency images, and employs a two-stage dynamic fusion framework to achieve feature fusion. This method can fully mine and integrate fault information from different perspectives, effectively overcoming the limitations of traditional methods that rely on a single perspective, and helping to improve the prediction accuracy of the fault diagnosis model. Furthermore, since the traditional binary cross-entropy loss function imposes a high gradient penalty on misclassified samples during training, it is prone to overfitting. This invention utilizes a binary cross-entropy loss function based on an adaptive adjustment mechanism, which, while ensuring the model's ability to learn from misclassified samples, enhances its robustness to noise, thus reducing the probability of overfitting. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This invention provides an intelligent diagnostic method for composite bearing faults based on dynamic fusion of multi-view features;

[0079] Figure 2 A schematic diagram of the framework of the bearing composite fault diagnosis method based on dynamic fusion of multi-view features provided in the embodiments of the present invention;

[0080] Figure 3 This is a schematic diagram illustrating the composition of composite fault types in bearing data provided in an embodiment of the present invention;

[0081] Figure 4 This is a schematic diagram of the CNN-BiLSTM-Attention-KAN model structure provided in an embodiment of the present invention;

[0082] Figure 5 This is a schematic diagram of the ResNet-KAN model structure provided in an embodiment of the present invention;

[0083] Figure 6 The t-SNE plot shows the prediction results of the bearing composite fault diagnosis method provided in this embodiment of the invention on an experimental dataset. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0085] The following will refer to the appendix. Figure 1 and Figure 2 Specific embodiments of the present invention are described in detail below. While specific embodiments of the invention have been discussed, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, the embodiments are provided so that this invention will be thorough and complete, and will fully convey the information of the invention to those skilled in the art.

[0086] This invention provides an intelligent diagnosis method for composite bearing faults based on dynamic fusion of multi-view features, the implementation process of which is shown in the attached figure. Figure 1 and Figure 2 As shown, it includes:

[0087] S100: Obtain raw data on composite bearing faults and construct a multi-label dataset based on bearing fault types.

[0088] The research data comes from the publicly available bearing dataset from the University of Paderborn, USA. This dataset has a vibration sampling frequency of 64 kHz and includes four fault types: normal, single fault, recurring fault, and compound fault. The experimental dataset consists of a mixture of single and compound fault data, resulting in four fault types based on the combination of fault location and fault mode: internal fatigue (IF), external fatigue (OF), internal pitting (IP), and external pitting (OP). The final dataset includes two single fault states (IF, OF), three compound fault states (IF-OF, IF-OP, IP-OP), and one healthy state (N), for a total of six bearing operating states. A sliding window sampling method was used to process the bearing vibration signals, with a window length of 512. 500 samples were collected for each state, for a total of 3000 samples. Each sample was encoded using a 5-dimensional multi-label binary vector. The label encoding order for each sample was [N, IF, OF, IP, OP]. The research data was divided in a 6:2:2 ratio: 60% of the data was used as the training set to train the fault diagnosis model, 20% of the data was used as the validation set to evaluate the generalization ability of the model during training, and 20% of the data was used as the test set to verify the performance of the bearing composite fault diagnosis model proposed in this invention. Figure 3 This is a schematic diagram illustrating the composition of composite fault types in bearing data provided in an embodiment of the present invention.

[0089] S200: Through Fast Fourier Transform and Continuous Wavelet Transform, the original data is converted into frequency domain features and time-frequency images, respectively, to achieve multi-view feature representation of the original data.

[0090] In the multi-view feature representation module, the original data is converted into frequency domain data and time-frequency image according to formulas (1) to (2).

[0091] S300: CNN-BiLSTM-Attention-KAN and ResNet-KAN models were constructed for frequency domain data and time-frequency images, respectively, to extract frequency domain features and time-frequency features.

[0092] In the frequency domain data feature extraction model, the input sequence length is 512. The kernel size of the first convolutional layer of the convolutional neural network is 32, the number of output channels is 32, the stride is 8, and max pooling is performed. In the second convolutional layer, the kernel size is 3, the number of output channels is 128, the stride is 1, and max pooling is performed again. The number of hidden units in BiLSTM is 256, the dimension of the attention mechanism in Attention is 512, the dimension of the first two hidden layers of KAN is 128, and the dimension of the last layer is 5. The CNN-BiLSTM-Attention-KAN model is constructed according to formulas (3), (5) to (10) to extract frequency domain features. In the time-frequency image feature extraction model, the input image dimension is 214×214×3, the number of modules in ResNet is 1, the dimension of the first two hidden layers of KAN is 128, and the dimension of the last layer is 5. The ResNet-KAN model is constructed according to formulas (3), (4) and (10) to extract time-frequency features. Figure 4 and Figure 5 The diagrams are schematics of the CNN-BiLSTM-Attention-KAN and ResNet-KAN models provided in the embodiments of the present invention.

[0093] S400: For the extracted frequency domain features and time-frequency features, a two-stage dynamic fusion framework is used for feature fusion. In the first stage, dynamic weights are generated through uncertainty perception and the model is trained using a binary cross-entropy loss function based on an adaptive adjustment mechanism. In the second stage, adaptive learning is performed through a regression network to further optimize the weight allocation, so as to improve the quality of the fused features and the classification performance in the fault diagnosis model.

[0094] Feature extraction model for frequency domain data Time-frequency image feature extraction model Extracted frequency domain features and time-frequency characteristics A two-stage dynamic fusion framework combining an uncertainty-aware strategy and an adaptive learning strategy is used to achieve effective fusion of multi-view features. In the first stage of this framework, the uncertainty-aware method is first used to calculate the uncertainty of each viewpoint, thereby obtaining the fusion weights. and Then, these weights are used to generate a fused feature vector. Finally, by Input to activation function In this process, the prediction probability for the first stage can be obtained. Specifically, in the first stage, a binary cross-entropy loss function based on an adaptive adjustment mechanism is introduced, and hyperparameters are configured differently for the classification difficulty of the samples: for easy-to-classify samples and difficult-to-classify samples, the hyperparameters of the loss function are adjusted accordingly. and The values ​​were set to 1 and 0.1 respectively, thus enhancing the model's robustness to noise while ensuring it learns difficult-to-classify samples. In the second stage, the optimized... and As input to the regression network, this method generates weights for the two views respectively. and After the second stage of fusion, a fusion vector can be obtained. Then input it into In order to obtain the final predicted probability for each category. The second stage uses a multilayer perceptron (MLP) as the regression network, with the first two hidden layers having a dimension of 128 and the last layer having an output dimension of 5, for adaptive learning of the optimal dynamic fusion weights. Through the above mechanism, combined with the mathematical models described in formulas (11) to (27), dynamic fusion of features from two perspectives can be achieved.

[0095] S500: Trainable parameters for optimizing fault diagnosis models based on the Adam algorithm.

[0096] Exponential weighted average of gradient attenuation coefficient =0.9, the exponentially weighted average of the squared gradients attenuation coefficient =0.999, constant =10⁻⁸, learning rate =0.001. The trainable parameters of the fault diagnosis model can be updated according to formulas (27), (28), and (29).

[0097] S600: Uses the trained model to perform fault diagnosis and obtain the final prediction result.

[0098] The trained fault diagnosis model was used on the test set to make predictions and obtain the final prediction results, so as to verify the performance of the fault diagnosis method proposed in this invention. Figure 6 The t-SNE plot shows the prediction results of the bearing composite fault diagnosis method provided in this embodiment of the invention on an experimental dataset.

[0099] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for intelligent diagnosis of composite bearing faults based on dynamic fusion of multi-view features, characterized in that, Includes the following steps: S100: Obtain raw data on composite bearing faults and construct a multi-label dataset based on bearing fault types; S200: Through Fast Fourier Transform and Continuous Wavelet Transform, the original data is converted into frequency domain data and time-frequency image respectively, so as to realize the multi-view feature representation of the original data; S300: CNN-BiLSTM-Attention-KAN and ResNet-KAN models were constructed for frequency domain data and time-frequency images, respectively, to extract frequency domain features and time-frequency features; S400: For the extracted frequency domain features and time-frequency features, a two-stage dynamic fusion framework is used for feature fusion. In the first stage, dynamic weights are generated through uncertainty perception and the model is trained using a binary cross-entropy loss function based on an adaptive adjustment mechanism. In the second stage, adaptive learning is performed through a regression network to further optimize the weight allocation, so as to improve the quality of the fused features and classification performance in the fault diagnosis model. S5 00: Optimize the trainable parameters of the fault diagnosis model based on the Adam algorithm; S600: Uses the trained model to perform fault diagnosis and obtain the final prediction result.

2. The intelligent diagnosis method for composite bearing faults based on dynamic fusion of multi-view features according to claim 1, characterized in that, The calculation process of the Fast Fourier Transform described in S200 specifically adopts the following formula: in, It is a discrete-time signal Length, It is the first signal in the frequency domain One frequency component, These are the basis functions of the discrete Fourier transform used to adjust the phase of the frequency components. It is the Fast Fourier Transform function; The calculation process of the continuous wavelet transform described in S200 specifically adopts the following formula: in, It is the original signal. Morley wavelets are used as basis functions. It is dimensionless time. It is a scaling factor. It is the translation factor.

3. The intelligent diagnosis method for bearing composite faults based on dynamic fusion of multi-view features according to claim 1, characterized in that, In step S300, the feature extraction process of the frequency domain data is as follows: First, it is divided into samples of length 512 and input into the convolutional layer of the convolutional neural network; second, the output of the convolutional layer is subjected to max pooling operation; third, the calculation result of the max pooling operation is input into the BiLSTM layer. Then, the calculation results of the BiLSTM layer are input into the Attention layer; Finally, the frequency domain features extracted by the Attention layer are input into the KAN layer to obtain fault classification prediction results based on frequency domain features.

4. The intelligent diagnosis method for bearing composite faults based on dynamic fusion of multi-view features according to claim 3, characterized in that, The convolution operation of the convolutional neural network described in S300 specifically adopts the following formula: in, yes The first layer One element, It is the first Convolutional regions of feature maps, It is a weight matrix. It's a bias. It is an activation function; The average pooling operation of the convolutional neural network described in S300 specifically uses the following formula: The max pooling operation of the convolutional neural network described in S300 specifically uses the following formula: in, These are the corresponding elements of the output feature map. For window size, Step size, and Both are indexes. It is the corresponding element in the input; In S300, when the results of the max pooling operation are input into the BiLSTM layer, the output of the forward hidden state of the BiLSTM model specifically uses the following formula: in, This refers to the LSTM computation process. The input weight matrix for the forward LSTM is... Here is the hidden state weight matrix of the forward LSTM. yes The input vector at time t, For forward LSTM in The hidden state at all times For forward LSTM in The hidden state at any given moment; The back hidden state output of the BiLSTM model is specifically calculated using the following formula: in, This refers to the LSTM computation process. The input weight matrix for the backward LSTM The hidden state weight matrix of the backward LSTM. yes The input vector at time t, For backward LSTM The hidden state at all times For backward LSTM in The hidden state at any given moment; The final output of the BiLSTM model is specifically expressed by the following formula: in, It is the fusion weight of the forward hidden state. It is the fusion weight of the backward hidden state. It is an activation function; In S300, the self-attention mechanism is calculated using the following formula when inputting the frequency domain features extracted by the Attention layer into the KAN layer: in, , It is the sequence length. It is the embedded dimension. , and The query, key, and value are obtained sequentially by performing a linear transformation on the input sequence. , and In order, they are the weight matrices for trainable queries, keys, and values. It is the result of calculation using the self-attention mechanism; In S300, the calculation process of the KAN model specifically adopts the following formula: in, It is the dimension of the input vector. It is the input vector. It is a one-dimensional continuous function. It is a one-dimensional continuous function within the closed interval [0,1]. and They are from 0 to Outer sum index from 1 to The inner summation index.

5. The intelligent diagnosis method for composite bearing faults based on dynamic fusion of multi-view features according to claim 1, characterized in that, In step S300, the feature extraction process of the time-frequency image is as follows: First, the dimensions of the time-frequency image are unified to 214×214×3; then, the time-frequency image and its fault label are input into the ResNet model to extract time-frequency features; finally, the extracted time-frequency features are input into the KAN layer to obtain the fault classification prediction result based on the time-frequency features.

6. The intelligent diagnosis method for bearing composite faults based on dynamic fusion of multi-view features according to claim 1, characterized in that, In the two-stage dynamic fusion framework described in S400, the dynamic weighted fusion specifically adopts the following formula: in, It is a multi-view classification model. It is the first Feature extraction model from multiple perspectives and The learnable parameters of the multi-view classification model are, in order, the first and second. Learnable parameters of a feature extraction model from multiple perspectives It is the first The fusion weights of features from each perspective , It is the number of viewpoint features. No. Input feature vectors from multiple perspectives; The weight design principle of the two-stage dynamic fusion framework described in S400 specifically adopts the following formula: in, It is the first The classification accuracy metric of the feature extraction model from the nth perspective, when the nth perspective is... In a feature extraction model based on a single perspective, the output is 1 when a sample is correctly classified, and a smaller value when it is misclassified. , and The order is the first The predicted and actual labels of samples in a feature extraction model from multiple perspectives; The theoretical design principles of the two-stage dynamic fusion framework described in S400 specifically adopt the following formula: in, It is the accuracy of multi-view classification. It is the Pearson correlation coefficient; In the first stage of the two-stage dynamic fusion framework described in S400, the traditional confidence ratio in the dynamic uncertainty perception of the first stage specifically adopts the following formula: in, It is the predicted probability of TOP-1. It is the predicted probability of TOP-2; In the first stage of the two-stage dynamic fusion framework described in S400, the evidence vector generated by evidence deep learning in the dynamic uncertainty perception of the two-stage dynamic fusion framework specifically adopts the following formula: in, It is the first The evidence vector of the feature extraction model from each perspective, where Softplus is the activation function; In the first stage of the two-stage dynamic fusion framework described in S400, the predicted probability output for each category in the dynamic perception of uncertainty is specifically calculated using the following formula: in, It is the first A feature extraction model from a single perspective for categories The predicted probability, , It is the first A feature extraction model from a single perspective for categories The Dirichlet distribution parameters, It is the first A feature extraction model from a single perspective for categories The vector of evidence, ,make ,but , It is the dimension of the labels in the sample; In the first stage of the two-stage dynamic fusion framework described in S400, the uncertainty dynamic perception output based on evidence is calculated using the following formula: in, and The order is the first Predicted probabilities of TOP-1 and TOP-2 categories in a feature extraction model from multiple perspectives and The order is the first Dirichlet distribution parameters of the top-1 and top-2 classes in the feature extraction model from each perspective and The order is the first Evidence vectors of the top-1 and top-2 categories in the feature extraction model from each perspective It is the first Predictive uncertainty of feature extraction models from multiple perspectives; In the first stage of the two-stage dynamic fusion framework described in S400, the output of the dynamic fusion weights in the uncertainty dynamic perception is specifically calculated using the following formula: in, It is the first Predictive uncertainty of feature extraction models from multiple perspectives The first Predictive determinism of a feature extraction model from multiple perspectives Is assigned to the first Dynamic fusion weights of features from multiple perspectives; In the first stage of the two-stage dynamic fusion framework described in S400, the uncertainty dynamic perception, the final classification evidence vector output is specifically calculated using the following formula: in, It is a category The Middle Evidence vectors from a feature extraction model based on individual perspectives. It is a category The vector of evidence; In the first stage of the two-stage dynamic fusion framework described in S400, the prediction probability output in the dynamic perception of uncertainty is specifically calculated using the following formula: in, It is a category The predicted probability; In the first stage of the two-stage dynamic fusion framework described in S400, the binary cross-entropy loss function based on the adaptive adjustment mechanism is specifically calculated using the following formula: in, It is the sample label dimension. It is the first The first sample The real label of each tag It is the first The first sample Predicted labels for each label. It is the first The first sample The prediction confidence of each label, It is the first The first sample The predicted probability of each label. It is a hyperparameter that determines whether a sample is easily classified when the model's prediction is correct. Set as When the model's prediction is incorrect, the sample is considered a hard-to-classify sample. Set as ; In the first stage of the two-stage dynamic fusion framework described in S400, the overall objective function is specifically calculated using the following formula: in, Input data The classification results of the multi-view classification model. It is the first Classification results of a feature extraction model from multiple perspectives; In the second stage of the two-stage dynamic fusion framework described in S400, a regression network is used to generate optimal weights and then normalize them, specifically using the following calculation formula: in, It is The first one obtained after normalization The fusion weights of features from each perspective It's about returning to the internet; In the second stage of the two-stage dynamic fusion framework described in S400, the regression-based multi-view dynamic fusion specifically adopts the following calculation formula: in, It is the second stage of the regression network. Dynamic fusion weights of features from various perspectives It is the first stage The learnable parameters of the feature extraction model from each perspective are frozen in the second stage. In the second stage of the two-stage dynamic fusion framework described in S400, the objective function is specifically calculated using the following formula:

7. The intelligent diagnosis method for composite bearing faults based on dynamic fusion of multi-view features according to claim 1, characterized in that, In S500, the Adam algorithm is used to update the trainable parameters of the diagnostic model, specifically using the following calculation formula: in, It is the parameter at time t+1. It is a parameter at time t. It's the learning rate. It is a constant with a very small value. It is the gradient The exponentially weighted average of the squares of, It is the gradient The exponentially weighted average, and It is the weighted average of two exponents. and The attenuation coefficient.

Images